US20060095729A1 - Multithreaded processor with multiple concurrent pipelines per thread - Google Patents

Multithreaded processor with multiple concurrent pipelines per thread Download PDF

Info

Publication number
US20060095729A1
US20060095729A1 US11/096,917 US9691705A US2006095729A1 US 20060095729 A1 US20060095729 A1 US 20060095729A1 US 9691705 A US9691705 A US 9691705A US 2006095729 A1 US2006095729 A1 US 2006095729A1
Authority
US
United States
Prior art keywords
instruction
instructions
processor
threads
multithreaded processor
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
US11/096,917
Other versions
US8074051B2 (en
Inventor
Erdem Hokenek
Mayan Moudgill
Michael Schulte
C. Glossner
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Qualcomm Inc
Original Assignee
Individual
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority to US11/096,917 priority Critical patent/US8074051B2/en
Application filed by Individual filed Critical Individual
Priority to KR1020127022421A priority patent/KR101303119B1/en
Priority to EP11001888A priority patent/EP2339455A3/en
Priority to PCT/US2005/011614 priority patent/WO2005101221A1/en
Priority to JP2007507460A priority patent/JP2007533007A/en
Priority to EP05732166A priority patent/EP1741033A4/en
Priority to EP11001890A priority patent/EP2339457A3/en
Priority to KR1020127022422A priority patent/KR101295569B1/en
Priority to EP11001889A priority patent/EP2339456A3/en
Priority to KR1020067022996A priority patent/KR101253549B1/en
Priority to KR1020127022423A priority patent/KR101236396B1/en
Priority to TW094111009A priority patent/TW200604940A/en
Assigned to SANDBRIDGE TECHNOLOGIES, INC. reassignment SANDBRIDGE TECHNOLOGIES, INC. ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS). Assignors: SCHULTE, MICHAEL J., GLOSSNER, C. JOHN, HOKENEK, ERDEM, MOUDGILL, MAYAN
Publication of US20060095729A1 publication Critical patent/US20060095729A1/en
Priority to US12/579,867 priority patent/US8959315B2/en
Priority to US12/579,912 priority patent/US8918627B2/en
Priority to US12/579,893 priority patent/US8892849B2/en
Assigned to ASPEN ACQUISITION CORPORATION reassignment ASPEN ACQUISITION CORPORATION ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS). Assignors: SANDBRIDGE TECHNOLOGIES, INC.
Priority to US13/282,800 priority patent/US8762688B2/en
Publication of US8074051B2 publication Critical patent/US8074051B2/en
Application granted granted Critical
Assigned to QUALCOMM INCORPORATED reassignment QUALCOMM INCORPORATED ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS). Assignors: ASPEN ACQUISITION CORPORATION
Active legal-status Critical Current
Adjusted expiration legal-status Critical

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • G06F9/3836Instruction issuing, e.g. dynamic instruction scheduling or out of order instruction execution
    • G06F9/3851Instruction issuing, e.g. dynamic instruction scheduling or out of order instruction execution from multiple instruction streams, e.g. multistreaming
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F15/00Digital computers in general; Data processing equipment in general
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/30098Register arrangements
    • G06F9/30101Special purpose registers
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/30098Register arrangements
    • G06F9/3012Organisation of register space, e.g. banked or distributed register file
    • G06F9/30123Organisation of register space, e.g. banked or distributed register file according to context, e.g. thread buffers
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/30003Arrangements for executing specific machine instructions
    • G06F9/30007Arrangements for executing specific machine instructions to perform operations on data operands
    • G06F9/30036Instructions to perform operations on packed data, e.g. vector, tile or matrix operations
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/30Arrangements for executing machine instructions, e.g. instruction decode
    • G06F9/38Concurrent instruction execution, e.g. pipeline, look ahead
    • G06F9/3885Concurrent instruction execution, e.g. pipeline, look ahead using a plurality of independent parallel functional units
    • G06F9/3893Concurrent instruction execution, e.g. pipeline, look ahead using a plurality of independent parallel functional units controlled in tandem, e.g. multiplier-accumulator

Definitions

  • the present invention relates generally to the field of digital data processors, and more particularly to multithreading and pipelining techniques for use in a digital signal processor (DSP) or other type of digital data processor.
  • DSP digital signal processor
  • Pipelining is a well-known processor implementation technique whereby multiple instructions are overlapped in execution.
  • Conventional pipelining techniques are described in, for example, John L. Hennessy and David A. Patterson, “Computer Architecture: A Quantitative Approach,” Third Edition, Morgan Kaufmann Publishers, Inc., San Francisco, Calif., 2003.
  • FIG. 1A shows an example involving the execution of two instructions without any overlap.
  • the two instructions are an integer add instruction addi r 0 , r 2 , 8 , and an integer multiplication instruction muli r 8 , r 3 , 4 .
  • the first instruction, addi performs an addition of the contents of register r 2 and an immediate value 8 , and stores the result in register r 0 .
  • each of the instructions includes the same four pipeline stages, denoted instruction fetch (IF), read (RD), execute (EX) and writeback (WB).
  • FIG. 1B shows the same two instructions but depicts how they may be overlapped using a conventional pipelining technique.
  • Each of the pipeline stages (IF, RD, EX and WB) is generally executed on a clock boundary.
  • the second instruction, muli may be started on the second clock cycle without requiring additional hardware.
  • the hardware associated with the IF, RD, EX and WB stages are shared between the two instructions, but the stages of one instruction are shifted in time relative to those of the other.
  • FIG. 2 illustrates a complication that may arise in a pipeline implementation.
  • the muli instruction requires as an operand the contents of register r 0 , and thus cannot read r 0 until the addi instruction has computed and written back the result of the addition operation to r 0 .
  • Processing of the muli instruction begins on the next clock cycle following the start of the addi instruction, but this process must stall and wait for the execution and writeback stages of the addi instruction to complete.
  • the empty cycles the muli instruction must wait for its operands to become available are typically called “bubbles” in the pipeline.
  • bypassing In single-threaded processors, a common method for reducing pipeline bubbles is known as bypassing, whereby instead of writing the computed value back to the register file in the WB stage, the result is forwarded directly to the processor execution unit that requires it. This reduces but does not eliminate bubbles in deeply pipelined machines. Also, it generally requires dependency checking and bypassing hardware, which unduly increases processor cost and complexity.
  • Multithreaded processors are processors that support simultaneous execution of multiple distinct instruction sequences or “threads.”
  • Conventional threading techniques are described in, for example, M. J. Flynn, “Computer Architecture: Pipelined and Parallel Processor Design,” Jones and Bartlett Publishers, Boston, Mass., 1995, and G. A. Blaauw and Frederick P. Brooks, “Computer Architecture: Concepts and Evolution,” Addison-Wesley, Reading, Mass., 1997, both of which are incorporated by reference herein.
  • the present invention in an illustrative embodiment provides a multithreaded processor which advantageously allows multiple concurrent pipelines per thread, and also supports pipeline shifting.
  • a multithreaded processor comprises a plurality of hardware thread units, an instruction decoder coupled to the thread units for decoding instructions received therefrom, and a plurality of execution units for executing the decoded instructions.
  • the multithreaded processor is configured for controlling an instruction issuance sequence for threads associated with respective ones of the hardware thread units. On a given processor clock cycle, only a designated one of the threads is permitted to issue one or more instructions, but the designated thread that is permitted to issue instructions varies over a plurality of clock cycles in accordance with the instruction issuance sequence.
  • the instructions are pipelined in a manner which permits at least a given one of the threads to support multiple concurrent instruction pipelines.
  • the instruction issuance sequence is determined using a token triggered threading approach. More specifically, in an arrangement in which the processor supports N threads, over a sequence of N consecutive processor clock cycles each of the N threads is permitted to issue instructions on only a corresponding one of the N consecutive processor clock cycles.
  • the illustrative embodiment allows each of the threads to issue up to three instructions on its corresponding one of the processor clock cycles.
  • the instructions are pipelined such that at least five separate instruction pipelines can be concurrently executing for different ones of the threads.
  • the pipelined instructions in the illustrative embodiment include a load/store instruction, an arithmetic logic unit instruction, an integer multiplication instruction, a vector multiplication instruction, and a vector multiplication and reduction instruction.
  • the vector multiplication and reduction instruction is pipelined using a number of stages which is greater than a total number of threads of the processor.
  • the vector multiplication and reduction instruction may comprise a pipeline with at least eleven stages, including an instruction decode stage, a vector register file read stage, at least two multiply stages, at least two add stages, an accumulator read stage, a plurality of reduction stages, and an accumulator writeback stage.
  • the accumulator read stage may be combined with another of the stages, such as an add stage.
  • Pipelines for respective vector multiplication and reduction instructions may be shifted relative to one another by a plurality of pipeline stages.
  • the present invention in the illustrative embodiment provides a number of significant advantages over conventional techniques. For example, a higher degree of concurrency is provided than that achievable using conventional approaches. Also, the need for dependency checking and bypassing hardware is eliminated, since computation results are guaranteed to be written back to the appropriate register file before they are needed by the next instruction from the same thread. Furthermore, the techniques help to limit processor power consumption.
  • FIGS. 1A and 1B show examples of instruction execution using conventional approaches.
  • FIG. 2 illustrates the manner in which stalls can occur in a conventional processor pipeline.
  • FIG. 3 shows an example of a pipeline of a multithreaded pipelined processor in an embodiment of the invention.
  • FIG. 4 shows an example of a multithreaded processor in which the present invention may be implemented.
  • FIG. 5 is a diagram illustrating an example token triggered multithreading approach that may be utilized in an embodiment of the invention.
  • FIG. 6 shows a number of example pipelines in an embodiment of the invention.
  • FIG. 7 illustrates the manner in which pipelines can be shifted to permit computation cycles which are longer than issue cycles, in accordance with one embodiment of the invention.
  • a given processor as described herein may be implemented in the form of one or more integrated circuits.
  • the present invention in an illustrative embodiment provides a pipelining technique suitable for use in a multithreaded processor. With this technique, multiple instructions from multiple threads can be concurrently executed in an efficient manner. As will be described in greater detail below, the illustrative embodiment uses variable length execution pipelines, staggered execution, and rotated start execution, to provide concurrent execution while maintaining low power operation. The illustrative embodiment provides a higher degree of concurrency than that achievable using conventional approaches.
  • FIG. 3 shows an example of a multithreaded pipeline that removes the “bubbles” that are introduced through pipelined execution.
  • the issued instructions each include the same four pipeline stages, namely, IF, RD, EX and WB. It is further assumed that there are three threads, and thus three hardware thread units or “contexts” issuing instructions in a sequential manner.
  • a given thread can generally be viewed in terms of hardware as well as software. The particular processor hardware associated with a given thread is therefore more particularly referred to herein as a hardware thread unit or simply a “context.”
  • an integer add instruction addi r 0 , r 2 , 8 is initially issued by a first one of the contexts on a first clock cycle.
  • the other two contexts issue instructions on respective subsequent clock cycles. It takes a total of three clock cycles for each of the contexts to issue an instruction.
  • the first context issues another instruction, namely an integer multiplication instruction muli r 8 , r 0 , 4 .
  • cycle 1 the IF stage of thread 1 is executed for the addi instruction.
  • cycle 2 the IF stage of thread 2 executes while at the same time the RD stage of thread 1 executes.
  • cycle 3 the IF stage of thread 3 executes, the RD stage of thread 2 executes, and the EX stage of thread 1 executes.
  • cycle 4 the IF stage of thread 1 of the muli instruction executes concurrently with the WB stage of the addi instruction. Simultaneously, the EX stage of thread 2 executes and the RD stage of thread 3 executes.
  • FIG. 3 example therefore serves to illustrate that with an appropriately-configured pipeline and a sufficient number of threads, all hardware contexts may be executing concurrently even though there is only a single instruction issued per context per cycle.
  • the particular number of threads and pipeline stages are purposes of illustration only, and not intended to reflect a preferred implementation. Those skilled in the art will be readily able to determine an appropriate number of threads and pipeline stages for a particular application given the teachings provided herein.
  • the present invention can be advantageously implemented in a multithreaded processor.
  • a more particular example of a multithreaded processor in which the invention may be implemented is described in U.S. patent application Ser. No. 10/269,372, filed Oct. 11, 2002 and entitled “Multithreaded Processor With Efficient Processing For Convergence Device Applications,” which is commonly assigned herewith and incorporated by reference herein.
  • This multithreaded processor may be configured to execute RISC-based control code, DSP code, Java code and network processing code. It includes a single instruction multiple data (SIMD) vector processing unit, a reduction unit, and long instruction word (LIW) compounded instruction execution.
  • SIMD single instruction multiple data
  • LIW long instruction word
  • the invention can be implemented in other multithreaded processors, or more generally other types of digital data processors. Another such processor will now be described with reference to FIG. 4 .
  • FIG. 4 shows an example of a multithreaded processor 400 incorporating a reduction unit 402 and an accumulator register file 406 .
  • the processor 400 is generally similar to that described in U.S. patent application Ser. No. 10/269,372, but incorporates reduction unit 402 and accumulator register file 406 configured as described in the above-cited U.S. patent application Ser. No. 10/841,261.
  • the multithreaded processor 400 includes, among other elements, a multithreaded cache memory 410 , a multithreaded data memory 412 , an instruction buffer 414 , an instruction decoder 416 , a register file 418 , and a memory management unit (MMU) 420 .
  • Each thread thus has a corresponding thread cache associated therewith in the multithreaded cache 410 .
  • the data memory 412 includes N distinct data memory instances, denoted data memories 412 - 1 , 412 - 2 , . . . 412 -N as shown.
  • the multithreaded cache 410 interfaces with a main memory (not shown) external to the processor 400 via the MMU 420 .
  • the MMU 420 like the cache 410 , includes a separate instance for the each of the N threads supported by the processor. The MMU 420 ensures that the appropriate instructions from main memory are loaded into the multithreaded cache 410 .
  • the data memory 412 is also typically directly connected to the above-noted external main memory, although this connection is also not explicitly shown in the figure. Also associated with the data memory 412 is a data buffer 430 .
  • the multithreaded cache 410 is used to store instructions to be executed by the multithreaded processor 400 , while the data memory 412 stores data that is operated on by the instructions. Instructions are fetched from the multithreaded cache 410 by the instruction decoder 416 and decoded. Depending upon the instruction type, the instruction decoder 416 may forward a given instruction or associated information to various other units within the processor, as will be described below.
  • the processor 400 includes a branch instruction queue (IQ) 440 and program counter (PC) registers 442 .
  • the program counter registers 442 include one instance for each of the threads.
  • the branch instruction queue 440 receives instructions from the instruction decoder 416 , and in conjunction with the program counter registers 442 provides input to an adder block 444 , which illustratively comprises a carry-propagate adder (CPA).
  • CPA carry-propagate adder
  • Elements 440 , 442 and 444 collectively comprise a branch unit of the processor 400 .
  • auxiliary registers may also be included in the processor 400 .
  • the register file 418 provides temporary storage of integer results. Instructions forwarded from the instruction decoder 416 to an integer instruction queue (IQ) 450 are decoded and the proper hardware thread unit is selected through the use of an offset unit 452 which is shown as including a separate instance for each of the threads.
  • the offset unit 452 inserts explicit bits into register file addresses so that independent thread data is not corrupted. For a given thread, these explicit bits may comprise, e.g., a corresponding thread identifier.
  • the register file 418 is coupled to input registers RA and RB, the outputs of which are coupled to an arithmetic logic unit (ALU) block 454 , which may comprise an adder.
  • ALU arithmetic logic unit
  • the input registers RA and RB are used in implementing instruction pipelining.
  • the output of the ALU block 454 is coupled to the data memory 412 .
  • the register file 418 , integer instruction queue 450 , offset unit 452 , elements RA and RB, and ALU block 454 collectively comprise an exemplary integer unit.
  • Instruction types executable in the processor 400 include Branch, Load, Store, Integer and Vector/SIMD instruction types. If a given instruction does not specify a Branch, Load, Store or Integer operation, it is a Vector/SIMD instruction. Other instruction types can also or alternatively be used.
  • the Integer and Vector/SIMD instruction types are examples of what are more generally referred to herein as integer and vector instruction types, respectively.
  • a vector IQ 456 receives Vector/SIMD instructions forwarded from the instruction decoder 416 .
  • a corresponding offset unit 458 shown as including a separate instance for each of the threads, serves to insert the appropriate bits to ensure that independent thread data is not corrupted.
  • a vector unit 460 of the processor 400 is separated into N distinct parallel portions, and includes a vector file 462 which is similarly divided.
  • the vector file 462 includes thirty-two registers, denoted VR 00 through VR 31 .
  • the vector file 462 serves substantially the same purpose as the register file 418 except that the former operates on Vector/SIMD instruction types.
  • the vector unit 460 illustratively comprises the vector instruction queue 456 , the offset unit 458 , the vector file 462 , and the arithmetic and storage elements associated therewith.
  • a Vector/SIMD block encoded either as a fractional or integer data type is read from the vector file 462 and is stored into architecturally visible registers VRA, VRB, VRC. From there, the flow proceeds through multipliers (MPY) that perform parallel concurrent multiplication of the Vector/SIMD data.
  • Adder units comprising carry-skip adders (CSAs) and CPAs may perform additional arithmetic operations. For example, one or more of the CSAs may be used to add in an accumulator value from a vector register file, and one or more of the CPAs may be used to perform a final addition for completion of a multiplication operation, as will be appreciated by those skilled in the art.
  • Computation results are stored in Result registers 464 , and are provided as input operands to the reduction unit 402 .
  • the reduction unit 402 sums the input operands in such a way that the summation result produced is the same as that which would be obtained if each operation were executed in series.
  • the reduced sum is stored in the accumulator register file 406 for further processing.
  • the MPY blocks When performing vector dot products, the MPY blocks perform four multiplies in parallel, the CSA and CPA units perform additional operations or simply pass along the multiplication results for storage in the Result registers 464 , and the reduction unit 402 sums the multiplication results, along with an accumulator value stored in the accumulator register file 406 . The result generated by the reduction unit is then stored in the accumulator register file for use in the next iteration, in the manner previously described.
  • the accumulator register file 406 in this example includes a total of sixteen accumulator registers denoted ACC 00 through ACC 15 .
  • the multithreaded processor 400 may make use of techniques for thread-based access to register files, as described in U.S. patent application Ser. No. 10/269,373, filed Oct. 11, 2002 and entitled “Method and Apparatus for Register File Port Reduction in a Multithreaded Processor,” which is commonly assigned herewith and incorporated by reference herein.
  • the multithreaded processor 400 is well suited for use in performing vector dot products and other types of parallel vector multiply and reduce operations, as described in the above-cited U.S. patent application Ser. No. 10/841,261.
  • the illustrative embodiment of the present invention utilizes an approach known as token triggered threading.
  • Token triggered threading is described in the above-cited U.S. patent application Ser. No. 10/269,245, now issued as U.S. Pat. No. 6,842,848.
  • the token triggered threading typically assigns different tokens to each of a plurality of threads of a multithreaded processor.
  • the token triggered threading may utilize a token to identify in association with a current processor clock cycle a particular one of the threads of the processor that will be permitted to issue an instruction for a subsequent clock cycle.
  • FIG. 5 shows an example of token triggered threading for an implementation of a multithreaded processor in which the number of threads N is eight. In general, all of the threads operate simultaneously, and each accesses a corresponding instance of the thread cache 110 and data memory 112 . As shown in FIG. 5 , the eight threads are denoted Thread 0 , Thread 1 , Thread 2 , . . . Thread 7 , and are illustrated as being serially interconnected in the form of a ring.
  • the tokens are arranged in a sequential or round-robin manner, such that the contexts will issue instructions sequentially.
  • tokens indicating the next context to issue an instruction may be arranged using other patterns, such as an alternating even-odd pattern.
  • other types of threading may be used in conjunction with the present invention.
  • token triggered threading is used in the illustrative embodiment, the invention does not require this particular type of multithreading, and other types of multithreading techniques can be used.
  • FIG. 6 illustrates the manner in which example instruction functions may be pipelined in the multithreaded processor 400 in accordance with the present invention.
  • this type of pipelining is preferably utilized in conjunction with the token triggered threading described previously, but it is to be appreciated that numerous other combinations of pipelining and threading may be used in implementing the invention.
  • the figure depicts example pipelines for Load/Store (Ld/St), Arithmetic Logic Unit (ALU), Integer Multiplication (I_Mul), Vector Multiplication (V_Mul), and Vector Multiplication and Reduction (V_Mul Reduce) instructions.
  • Ld/St Load/Store
  • ALU Arithmetic Logic Unit
  • I_Mul Integer Multiplication
  • V_Mul Vector Multiplication
  • V_Mul Reduce Vector Multiplication and Reduction
  • stage 0 Inst Dec
  • stage 1 RF Read
  • stage 2 Agen
  • stage 3 Xfer
  • stage 4 Int/Ext
  • a determination is made as to whether the memory access is to internal or external memory.
  • stages 5 - 7 Mem 0 , Mem 1 , Mem 2
  • the value is read from or written to memory.
  • stage 8 the value read from memory on a Load instruction is written into the register file.
  • stage 0 The ALU pipeline has seven stages, denoted stage 0 through stage 6 .
  • stage 0 Inst Dec
  • stage 1 a wait cycle is inserted. This allows the Ld/St and ALU hardware to share the same register file read ports.
  • stage 2 RF Read
  • the operands for the arithmetic function are read from the register file.
  • Stages 3 and 4 (Exec 1 , Exec 2 ) then compute the arithmetic result (e.g., an add, compare, shift, etc.).
  • stage 5 Xfer
  • stage 6 the result is written back into the register file.
  • the I_Mul pipeline is similar to the ALU pipeline, as they share common architected resources.
  • the figure indicates that the pipeline stages are identical except for an additional execution stage (Exec 3 ) in the I_Mul pipeline. Thus, an additional cycle is available for computing the result of a multiply.
  • stage 0 (Inst Dec) is as in all instructions and allows for routing of the decoded instruction to the correct pipeline.
  • stage 1 VRF Read
  • stage 2 - 5 (MPY 1 , MPY 2 , Add 1 , Add 2 ) perform the multi-element vector arithmetic.
  • the two add stages are present to convert the multiplication results from carry-save format back into two's complement format. Additionally, if the vectors only require simple arithmetic, this can be performed in the add stages.
  • stage 6 (Xfer) the results are transferred back to the vector register file, and in stage 7 (WB), the results are written back.
  • the V_Mul Reduce pipeline is similar to the V_Mul pipeline except that an additional reduction operation is performed.
  • the reduction takes the 4 vector element products, along with an accumulator operand, and reduces them to a single scalar element. Typically this involves adding all of the products to the accumulator or subtracting all of the products from the accumulator, although other combinations are possible.
  • the V_Mul and V_Mul Reduce pipelines are the same until stage 5 . In stage 5 (Add 2 , ACC Read), an additional architected accumulator register file is read. This value is arithmetically combined with the vector elements and reduced to a single scalar.
  • all five processor pipelines may be simultaneously active with instructions from multiple hardware thread units. This fills potential bubbles in the pipeline with work from other thread units.
  • V_Mul Reduce pipeline may be shifted in locality from a V_Mul pipeline in that the back-to-back reduction operations of the V_Mul Reduce pipeline do not cause bubbles. It appears that such a shift might lead to pipeline bubbles because the V_Mul Reduce pipeline is longer in duration than the number of hardware thread units (eight in this implementation). In other words, the computational cycle of the pipeline (eleven clock cycles for V_Mul Reduce) is longer than the issue cycle (each thread gets to issue once every eight clock cycles). In fact, this does not happen because the accumulator register file read phase is shifted from the V_Mul pipeline computations.
  • FIG. 7 illustrates the manner in which multiple V_Mul Reduce pipelines can be shifted relative to one another so as to permit computation cycles which are longer than issue cycles. Note that the figure starts from cycle 5 of the first V_Mul Reduce instruction. Since there are eight thread units in this implementation, the next V_Mul Reduce instruction will issue on cycle 8 . As can be seen in the figure, the accumulator register file is written back in cycle 10 . The operands, however, are not read by the second instruction until cycle 13 . The second V_Mul Reduce pipeline can be thought of as being shifted in locality from the first V_Mul Reduce pipeline. This allows the lengthening of execution phases without causing bubbles in the pipeline.
  • the illustrative embodiment described above advantageously allows multiple concurrent pipelines per thread and provides for pipeline shifting in deeply multithreaded pipelines. It also eliminates the need for dependency checking and bypassing hardware, since results are guaranteed to be written back to the register file before they are needed by the next instruction from the same thread.

Abstract

A multithreaded processor comprises a plurality of hardware thread units, an instruction decoder coupled to the thread units for decoding instructions received therefrom, and a plurality of execution units for executing the decoded instructions. The multithreaded processor is configured for controlling an instruction issuance sequence for threads associated with respective ones of the hardware thread units. On a given processor clock cycle, only a designated one of the threads is permitted to issue one or more instructions, but the designated thread that is permitted to issue instructions varies over a plurality of clock cycles in accordance with the instruction issuance sequence. The instructions are pipelined in a manner which permits at least a given one of the threads to support multiple concurrent instruction pipelines.

Description

    RELATED APPLICATION(S)
  • The present application claims the priority of U.S. Provisional Application Ser. No. 60/560,199, filed Apr. 7, 2004 and entitled “Processor Pipeline With Multithreaded Support,” which is incorporated by reference herein.
  • The present application is also related to U.S. patent application Ser. No. 10/841,261, filed May 7, 2004 and entitled “Processor Reduction Unit for Accumulation of Multiple Operands With or Without Saturation,” which is incorporated by reference herein.
  • FIELD OF THE INVENTION
  • The present invention relates generally to the field of digital data processors, and more particularly to multithreading and pipelining techniques for use in a digital signal processor (DSP) or other type of digital data processor.
  • BACKGROUND OF THE INVENTION
  • Pipelining is a well-known processor implementation technique whereby multiple instructions are overlapped in execution. Conventional pipelining techniques are described in, for example, John L. Hennessy and David A. Patterson, “Computer Architecture: A Quantitative Approach,” Third Edition, Morgan Kaufmann Publishers, Inc., San Francisco, Calif., 2003.
  • FIG. 1A shows an example involving the execution of two instructions without any overlap. In this example, the two instructions are an integer add instruction addi r0, r2, 8, and an integer multiplication instruction muli r8, r3, 4. The first instruction, addi, performs an addition of the contents of register r2 and an immediate value 8, and stores the result in register r0. It is assumed for simplicity and clarity of illustration that each of the instructions includes the same four pipeline stages, denoted instruction fetch (IF), read (RD), execute (EX) and writeback (WB).
  • In the first stage (IF) instructions are fetched from memory and decoded. In the second stage (RD) the operands are read from the register file. In the third stage (EX) the addition is performed. Finally, in the fourth stage (WB) the results are written back into the register file at location r0. When the addi instruction has completed, the next instruction muli is started. The muli instruction performs an addition of the contents of register r3 and an immediate value 4, and stores the result in register r8.
  • FIG. 1B shows the same two instructions but depicts how they may be overlapped using a conventional pipelining technique. Each of the pipeline stages (IF, RD, EX and WB) is generally executed on a clock boundary. The second instruction, muli, may be started on the second clock cycle without requiring additional hardware. The hardware associated with the IF, RD, EX and WB stages are shared between the two instructions, but the stages of one instruction are shifted in time relative to those of the other.
  • FIG. 2 illustrates a complication that may arise in a pipeline implementation. In this example, the muli instruction requires as an operand the contents of register r0, and thus cannot read r0 until the addi instruction has computed and written back the result of the addition operation to r0. Processing of the muli instruction begins on the next clock cycle following the start of the addi instruction, but this process must stall and wait for the execution and writeback stages of the addi instruction to complete. The empty cycles the muli instruction must wait for its operands to become available are typically called “bubbles” in the pipeline.
  • In single-threaded processors, a common method for reducing pipeline bubbles is known as bypassing, whereby instead of writing the computed value back to the register file in the WB stage, the result is forwarded directly to the processor execution unit that requires it. This reduces but does not eliminate bubbles in deeply pipelined machines. Also, it generally requires dependency checking and bypassing hardware, which unduly increases processor cost and complexity.
  • It is also possible to reduce pipeline stalls through the use of multithreading. Multithreaded processors are processors that support simultaneous execution of multiple distinct instruction sequences or “threads.” Conventional threading techniques are described in, for example, M. J. Flynn, “Computer Architecture: Pipelined and Parallel Processor Design,” Jones and Bartlett Publishers, Boston, Mass., 1995, and G. A. Blaauw and Frederick P. Brooks, “Computer Architecture: Concepts and Evolution,” Addison-Wesley, Reading, Mass., 1997, both of which are incorporated by reference herein.
  • However, these and other conventional approaches generally do not allow multiple concurrent pipelines per thread, nor do they support pipeline shifting.
  • Accordingly, techniques are needed which can provide improved pipelining in a multithreaded digital data processor.
  • SUMMARY OF THE INVENTION
  • The present invention in an illustrative embodiment provides a multithreaded processor which advantageously allows multiple concurrent pipelines per thread, and also supports pipeline shifting.
  • In accordance with one aspect of the invention, a multithreaded processor comprises a plurality of hardware thread units, an instruction decoder coupled to the thread units for decoding instructions received therefrom, and a plurality of execution units for executing the decoded instructions. The multithreaded processor is configured for controlling an instruction issuance sequence for threads associated with respective ones of the hardware thread units. On a given processor clock cycle, only a designated one of the threads is permitted to issue one or more instructions, but the designated thread that is permitted to issue instructions varies over a plurality of clock cycles in accordance with the instruction issuance sequence. The instructions are pipelined in a manner which permits at least a given one of the threads to support multiple concurrent instruction pipelines.
  • In the illustrative embodiment, the instruction issuance sequence is determined using a token triggered threading approach. More specifically, in an arrangement in which the processor supports N threads, over a sequence of N consecutive processor clock cycles each of the N threads is permitted to issue instructions on only a corresponding one of the N consecutive processor clock cycles.
  • The illustrative embodiment allows each of the threads to issue up to three instructions on its corresponding one of the processor clock cycles. The instructions are pipelined such that at least five separate instruction pipelines can be concurrently executing for different ones of the threads.
  • The pipelined instructions in the illustrative embodiment include a load/store instruction, an arithmetic logic unit instruction, an integer multiplication instruction, a vector multiplication instruction, and a vector multiplication and reduction instruction.
  • In accordance with another aspect of the invention, the vector multiplication and reduction instruction is pipelined using a number of stages which is greater than a total number of threads of the processor. For example, the vector multiplication and reduction instruction may comprise a pipeline with at least eleven stages, including an instruction decode stage, a vector register file read stage, at least two multiply stages, at least two add stages, an accumulator read stage, a plurality of reduction stages, and an accumulator writeback stage. The accumulator read stage may be combined with another of the stages, such as an add stage. Pipelines for respective vector multiplication and reduction instructions may be shifted relative to one another by a plurality of pipeline stages.
  • The present invention in the illustrative embodiment provides a number of significant advantages over conventional techniques. For example, a higher degree of concurrency is provided than that achievable using conventional approaches. Also, the need for dependency checking and bypassing hardware is eliminated, since computation results are guaranteed to be written back to the appropriate register file before they are needed by the next instruction from the same thread. Furthermore, the techniques help to limit processor power consumption.
  • BRIEF DESCRIPTION OF THE INVENTION
  • FIGS. 1A and 1B show examples of instruction execution using conventional approaches.
  • FIG. 2 illustrates the manner in which stalls can occur in a conventional processor pipeline.
  • FIG. 3 shows an example of a pipeline of a multithreaded pipelined processor in an embodiment of the invention.
  • FIG. 4 shows an example of a multithreaded processor in which the present invention may be implemented.
  • FIG. 5 is a diagram illustrating an example token triggered multithreading approach that may be utilized in an embodiment of the invention.
  • FIG. 6 shows a number of example pipelines in an embodiment of the invention.
  • FIG. 7 illustrates the manner in which pipelines can be shifted to permit computation cycles which are longer than issue cycles, in accordance with one embodiment of the invention.
  • DETAILED DESCRIPTION OF THE INVENTION
  • The present invention will be described in the context of an exemplary multithreaded processor. It should be understood, however, that the invention does not require the particular arrangements shown, and can be implemented using other types of digital data processors and associated processing circuitry.
  • A given processor as described herein may be implemented in the form of one or more integrated circuits.
  • The present invention in an illustrative embodiment provides a pipelining technique suitable for use in a multithreaded processor. With this technique, multiple instructions from multiple threads can be concurrently executed in an efficient manner. As will be described in greater detail below, the illustrative embodiment uses variable length execution pipelines, staggered execution, and rotated start execution, to provide concurrent execution while maintaining low power operation. The illustrative embodiment provides a higher degree of concurrency than that achievable using conventional approaches.
  • FIG. 3 shows an example of a multithreaded pipeline that removes the “bubbles” that are introduced through pipelined execution. As in the previous examples herein, it is assumed for simplicity and clarity of illustration in this example that the issued instructions each include the same four pipeline stages, namely, IF, RD, EX and WB. It is further assumed that there are three threads, and thus three hardware thread units or “contexts” issuing instructions in a sequential manner. In a typical implementation of a multithreaded processor, a given thread can generally be viewed in terms of hardware as well as software. The particular processor hardware associated with a given thread is therefore more particularly referred to herein as a hardware thread unit or simply a “context.”
  • In this example, an integer add instruction addi r0, r2, 8 is initially issued by a first one of the contexts on a first clock cycle. The other two contexts issue instructions on respective subsequent clock cycles. It takes a total of three clock cycles for each of the contexts to issue an instruction. On a fourth clock cycle, the first context issues another instruction, namely an integer multiplication instruction muli r8, r0, 4.
  • More specifically, in cycle 1, the IF stage of thread 1 is executed for the addi instruction. In cycle 2, the IF stage of thread 2 executes while at the same time the RD stage of thread 1 executes. In cycle 3, the IF stage of thread 3 executes, the RD stage of thread 2 executes, and the EX stage of thread 1 executes. In cycle 4, the IF stage of thread 1 of the muli instruction executes concurrently with the WB stage of the addi instruction. Simultaneously, the EX stage of thread 2 executes and the RD stage of thread 3 executes.
  • It can be seen from this example that multiple instructions from the same and different threads are overlapped and concurrently executing. It can also be seen that there are no bubbles in the pipeline even though the results of the addi instruction are required by the muli instruction. The FIG. 3 example therefore serves to illustrate that with an appropriately-configured pipeline and a sufficient number of threads, all hardware contexts may be executing concurrently even though there is only a single instruction issued per context per cycle. The particular number of threads and pipeline stages are purposes of illustration only, and not intended to reflect a preferred implementation. Those skilled in the art will be readily able to determine an appropriate number of threads and pipeline stages for a particular application given the teachings provided herein.
  • As indicated previously, the present invention can be advantageously implemented in a multithreaded processor. A more particular example of a multithreaded processor in which the invention may be implemented is described in U.S. patent application Ser. No. 10/269,372, filed Oct. 11, 2002 and entitled “Multithreaded Processor With Efficient Processing For Convergence Device Applications,” which is commonly assigned herewith and incorporated by reference herein. This multithreaded processor may be configured to execute RISC-based control code, DSP code, Java code and network processing code. It includes a single instruction multiple data (SIMD) vector processing unit, a reduction unit, and long instruction word (LIW) compounded instruction execution. Examples of threading and pipelining techniques suitable for use with this exemplary multithreaded processor are described in U.S. patent application Ser. No. 10/269,245, filed Oct. 11, 2002 and entitled “Method and Apparatus for Token Triggered Multithreading,” now issued as U.S. Pat. No. 6,842,848, which is commonly assigned herewith and incorporated by reference herein.
  • The invention can be implemented in other multithreaded processors, or more generally other types of digital data processors. Another such processor will now be described with reference to FIG. 4.
  • FIG. 4 shows an example of a multithreaded processor 400 incorporating a reduction unit 402 and an accumulator register file 406. The processor 400 is generally similar to that described in U.S. patent application Ser. No. 10/269,372, but incorporates reduction unit 402 and accumulator register file 406 configured as described in the above-cited U.S. patent application Ser. No. 10/841,261.
  • The multithreaded processor 400 includes, among other elements, a multithreaded cache memory 410, a multithreaded data memory 412, an instruction buffer 414, an instruction decoder 416, a register file 418, and a memory management unit (MMU) 420. The multithreaded cache 410 includes a plurality of thread caches 410-1, 410-2, . . . 410-N, where N generally denotes the number of threads supported by the multithreaded processor 400, and in this particular example is given by N=4. Of course, other values of N may be used, as will be readily apparent to those skilled in the art.
  • Each thread thus has a corresponding thread cache associated therewith in the multithreaded cache 410. Similarly, the data memory 412 includes N distinct data memory instances, denoted data memories 412-1, 412-2, . . . 412-N as shown.
  • The multithreaded cache 410 interfaces with a main memory (not shown) external to the processor 400 via the MMU 420. The MMU 420, like the cache 410, includes a separate instance for the each of the N threads supported by the processor. The MMU 420 ensures that the appropriate instructions from main memory are loaded into the multithreaded cache 410.
  • The data memory 412 is also typically directly connected to the above-noted external main memory, although this connection is also not explicitly shown in the figure. Also associated with the data memory 412 is a data buffer 430.
  • In general, the multithreaded cache 410 is used to store instructions to be executed by the multithreaded processor 400, while the data memory 412 stores data that is operated on by the instructions. Instructions are fetched from the multithreaded cache 410 by the instruction decoder 416 and decoded. Depending upon the instruction type, the instruction decoder 416 may forward a given instruction or associated information to various other units within the processor, as will be described below.
  • The processor 400 includes a branch instruction queue (IQ) 440 and program counter (PC) registers 442. The program counter registers 442 include one instance for each of the threads. The branch instruction queue 440 receives instructions from the instruction decoder 416, and in conjunction with the program counter registers 442 provides input to an adder block 444, which illustratively comprises a carry-propagate adder (CPA). Elements 440, 442 and 444 collectively comprise a branch unit of the processor 400. Although not shown in the figure, auxiliary registers may also be included in the processor 400.
  • The register file 418 provides temporary storage of integer results. Instructions forwarded from the instruction decoder 416 to an integer instruction queue (IQ) 450 are decoded and the proper hardware thread unit is selected through the use of an offset unit 452 which is shown as including a separate instance for each of the threads. The offset unit 452 inserts explicit bits into register file addresses so that independent thread data is not corrupted. For a given thread, these explicit bits may comprise, e.g., a corresponding thread identifier.
  • As shown in the figure, the register file 418 is coupled to input registers RA and RB, the outputs of which are coupled to an arithmetic logic unit (ALU) block 454, which may comprise an adder. The input registers RA and RB are used in implementing instruction pipelining. The output of the ALU block 454 is coupled to the data memory 412.
  • The register file 418, integer instruction queue 450, offset unit 452, elements RA and RB, and ALU block 454 collectively comprise an exemplary integer unit.
  • Instruction types executable in the processor 400 include Branch, Load, Store, Integer and Vector/SIMD instruction types. If a given instruction does not specify a Branch, Load, Store or Integer operation, it is a Vector/SIMD instruction. Other instruction types can also or alternatively be used. The Integer and Vector/SIMD instruction types are examples of what are more generally referred to herein as integer and vector instruction types, respectively.
  • A vector IQ 456 receives Vector/SIMD instructions forwarded from the instruction decoder 416. A corresponding offset unit 458, shown as including a separate instance for each of the threads, serves to insert the appropriate bits to ensure that independent thread data is not corrupted.
  • A vector unit 460 of the processor 400 is separated into N distinct parallel portions, and includes a vector file 462 which is similarly divided. The vector file 462 includes thirty-two registers, denoted VR00 through VR31. The vector file 462 serves substantially the same purpose as the register file 418 except that the former operates on Vector/SIMD instruction types.
  • The vector unit 460 illustratively comprises the vector instruction queue 456, the offset unit 458, the vector file 462, and the arithmetic and storage elements associated therewith.
  • The operation of the vector unit 460 is as follows. A Vector/SIMD block encoded either as a fractional or integer data type is read from the vector file 462 and is stored into architecturally visible registers VRA, VRB, VRC. From there, the flow proceeds through multipliers (MPY) that perform parallel concurrent multiplication of the Vector/SIMD data. Adder units comprising carry-skip adders (CSAs) and CPAs may perform additional arithmetic operations. For example, one or more of the CSAs may be used to add in an accumulator value from a vector register file, and one or more of the CPAs may be used to perform a final addition for completion of a multiplication operation, as will be appreciated by those skilled in the art. Computation results are stored in Result registers 464, and are provided as input operands to the reduction unit 402. The reduction unit 402 sums the input operands in such a way that the summation result produced is the same as that which would be obtained if each operation were executed in series. The reduced sum is stored in the accumulator register file 406 for further processing.
  • When performing vector dot products, the MPY blocks perform four multiplies in parallel, the CSA and CPA units perform additional operations or simply pass along the multiplication results for storage in the Result registers 464, and the reduction unit 402 sums the multiplication results, along with an accumulator value stored in the accumulator register file 406. The result generated by the reduction unit is then stored in the accumulator register file for use in the next iteration, in the manner previously described.
  • The accumulator register file 406 in this example includes a total of sixteen accumulator registers denoted ACC00 through ACC15.
  • The multithreaded processor 400 may make use of techniques for thread-based access to register files, as described in U.S. patent application Ser. No. 10/269,373, filed Oct. 11, 2002 and entitled “Method and Apparatus for Register File Port Reduction in a Multithreaded Processor,” which is commonly assigned herewith and incorporated by reference herein.
  • The multithreaded processor 400 is well suited for use in performing vector dot products and other types of parallel vector multiply and reduce operations, as described in the above-cited U.S. patent application Ser. No. 10/841,261.
  • The illustrative embodiment of the present invention utilizes an approach known as token triggered threading. Token triggered threading is described in the above-cited U.S. patent application Ser. No. 10/269,245, now issued as U.S. Pat. No. 6,842,848. The token triggered threading typically assigns different tokens to each of a plurality of threads of a multithreaded processor. For example, the token triggered threading may utilize a token to identify in association with a current processor clock cycle a particular one of the threads of the processor that will be permitted to issue an instruction for a subsequent clock cycle.
  • FIG. 5 shows an example of token triggered threading for an implementation of a multithreaded processor in which the number of threads N is eight. In general, all of the threads operate simultaneously, and each accesses a corresponding instance of the thread cache 110 and data memory 112. As shown in FIG. 5, the eight threads are denoted Thread 0, Thread 1, Thread 2, . . . Thread 7, and are illustrated as being serially interconnected in the form of a ring.
  • In accordance with the token triggered threading illustrated in FIG. 5, all of the hardware thread units or contexts are permitted to simultaneously execute instructions, but only one context may issue an instruction in a particular clock cycle of the processor. In other words, all contexts execute simultaneously but only one context is active on a particular clock cycle. Therefore, if there are a total of C contexts it will require C clock cycles to issue an instruction from all contexts. Each clock cycle, one of the contexts issues an instruction, and the next thread to issue an instruction is indicated by a token.
  • In the FIG. 5 example, the tokens are arranged in a sequential or round-robin manner, such that the contexts will issue instructions sequentially. However, tokens indicating the next context to issue an instruction may be arranged using other patterns, such as an alternating even-odd pattern. Also, as noted above, other types of threading may be used in conjunction with the present invention.
  • Although token triggered threading is used in the illustrative embodiment, the invention does not require this particular type of multithreading, and other types of multithreading techniques can be used.
  • FIG. 6 illustrates the manner in which example instruction functions may be pipelined in the multithreaded processor 400 in accordance with the present invention. In the illustrative embodiment of the invention, this type of pipelining is preferably utilized in conjunction with the token triggered threading described previously, but it is to be appreciated that numerous other combinations of pipelining and threading may be used in implementing the invention.
  • The figure depicts example pipelines for Load/Store (Ld/St), Arithmetic Logic Unit (ALU), Integer Multiplication (I_Mul), Vector Multiplication (V_Mul), and Vector Multiplication and Reduction (V_Mul Reduce) instructions. In this implementation, up to three pipelines may be simultaneously started and all five may be in various phases of execution concurrently.
  • The Ld/St pipeline has nine stages, denoted stage 0 through stage 8. In the first stage, stage 0 (Inst Dec), an instruction is fetched and decoded. This stage is common to all five pipelines and determines which queue the instructions should be routed to. In stage 1 (RF Read), the register file operands are read. This will form the base address for the load or store operation. In the case of a store instruction, the data to be stored is also read. In stage 2 (Agen), any immediate values are added to the address and the full address is generated. In stage 3 (Xfer), the computed address is transferred to the memory subsystem. In stage 4 (Int/Ext), a determination is made as to whether the memory access is to internal or external memory. In stages 5-7 (Mem0, Mem1, Mem2), the value is read from or written to memory. In stage 8 (WB), the value read from memory on a Load instruction is written into the register file.
  • The ALU pipeline has seven stages, denoted stage 0 through stage 6. As in the Ld/St pipeline, the first stage, stage 0 (Inst Dec), fetches and decodes all instructions. In stage 1 (Wait), a wait cycle is inserted. This allows the Ld/St and ALU hardware to share the same register file read ports. In the following stage, stage 2 (RF Read), the operands for the arithmetic function are read from the register file. Stages 3 and 4 (Exec1, Exec2) then compute the arithmetic result (e.g., an add, compare, shift, etc.). In stage 5 (Xfer), the result is transferred to the register file. In stage 6 (WB), the result is written back into the register file.
  • The I_Mul pipeline is similar to the ALU pipeline, as they share common architected resources. The figure indicates that the pipeline stages are identical except for an additional execution stage (Exec 3) in the I_Mul pipeline. Thus, an additional cycle is available for computing the result of a multiply.
  • The V_Mul pipeline uses different architected resources than the previously-described ALU and I_Mul pipelines. It may therefore execute concurrently with those instructions without resource conflicts. Stage 0 (Inst Dec) is as in all instructions and allows for routing of the decoded instruction to the correct pipeline. In stage 1 (VRF Read) the vector register file operands are read. Stages 2-5 (MPY1, MPY2, Add1, Add2) perform the multi-element vector arithmetic. The two add stages are present to convert the multiplication results from carry-save format back into two's complement format. Additionally, if the vectors only require simple arithmetic, this can be performed in the add stages. In stage 6 (Xfer), the results are transferred back to the vector register file, and in stage 7 (WB), the results are written back.
  • The V_Mul Reduce pipeline is similar to the V_Mul pipeline except that an additional reduction operation is performed. The reduction takes the 4 vector element products, along with an accumulator operand, and reduces them to a single scalar element. Typically this involves adding all of the products to the accumulator or subtracting all of the products from the accumulator, although other combinations are possible. The V_Mul and V_Mul Reduce pipelines are the same until stage 5. In stage 5 (Add2, ACC Read), an additional architected accumulator register file is read. This value is arithmetically combined with the vector elements and reduced to a single scalar. Four stages (Reduce1, Reduce2, Reduce3, Reduce4) are devoted to this reduction and then the scalar value is written back to the accumulator register file (i.e., a different architected space from the vector register file) in stage 10 (ACC WB).
  • If a single thread issued instructions each cycle as in FIG. 2, bubbles would be induced in the pipeline. However, just as in the simplified case shown in FIG. 3 where there is only one type of pipeline, in the illustrative embodiment all five of the processor pipelines are interleaved and multithreaded to avoid bubbles. Each hardware thread unit issues up to three instructions in accordance with token triggered threading as shown in FIG. 5. This ensures that no threads stall and all threads will complete without deadlock.
  • As mentioned previously, in this implementation, all five processor pipelines may be simultaneously active with instructions from multiple hardware thread units. This fills potential bubbles in the pipeline with work from other thread units.
  • It should be noted that a given V_Mul Reduce pipeline may be shifted in locality from a V_Mul pipeline in that the back-to-back reduction operations of the V_Mul Reduce pipeline do not cause bubbles. It appears that such a shift might lead to pipeline bubbles because the V_Mul Reduce pipeline is longer in duration than the number of hardware thread units (eight in this implementation). In other words, the computational cycle of the pipeline (eleven clock cycles for V_Mul Reduce) is longer than the issue cycle (each thread gets to issue once every eight clock cycles). In fact, this does not happen because the accumulator register file read phase is shifted from the V_Mul pipeline computations.
  • FIG. 7 illustrates the manner in which multiple V_Mul Reduce pipelines can be shifted relative to one another so as to permit computation cycles which are longer than issue cycles. Note that the figure starts from cycle 5 of the first V_Mul Reduce instruction. Since there are eight thread units in this implementation, the next V_Mul Reduce instruction will issue on cycle 8. As can be seen in the figure, the accumulator register file is written back in cycle 10. The operands, however, are not read by the second instruction until cycle 13. The second V_Mul Reduce pipeline can be thought of as being shifted in locality from the first V_Mul Reduce pipeline. This allows the lengthening of execution phases without causing bubbles in the pipeline.
  • The illustrative embodiment described above advantageously allows multiple concurrent pipelines per thread and provides for pipeline shifting in deeply multithreaded pipelines. It also eliminates the need for dependency checking and bypassing hardware, since results are guaranteed to be written back to the register file before they are needed by the next instruction from the same thread.
  • It should be noted that the particular processor, multithreading, pipelining and shifting arrangements shown in the figures are presented by way of illustrative example only, and additional or alternative elements not explicitly shown may be included, as will be apparent to those skilled in the art.
  • It should also be emphasized that the present invention does not require the particular multithreaded processor configuration shown in FIG. 4. The invention can be implemented in a wide variety of other multithreaded processor configurations.
  • Thus, the above-described embodiments of the invention are intended to be illustrative only, and numerous alternative embodiments within the scope of the appended claims will be apparent to those skilled in the art. For example, the particular arrangement of hardware thread units, instruction decoder and execution units shown in FIG. 4 may be altered in other embodiments, and the invention should not be construed as requiring any particular type or arrangement of such elements. Also, as noted above, pipeline configurations, threading types and instruction formats may be varied to accommodate the particular needs of a given application.

Claims (20)

1. A multithreaded processor comprising:
a plurality of hardware thread units;
an instruction decoder coupled to the thread units for decoding instructions received therefrom; and
a plurality of execution units for executing the decoded instructions;
wherein the multithreaded processor is configured for controlling an instruction issuance sequence for a plurality of threads associated with respective ones of the hardware thread units;
wherein on a given processor clock cycle only a designated one of the threads is permitted to issue one or more instructions, the designated thread that is permitted to issue instructions varying over a plurality of clock cycles in accordance with the instruction issuance sequence; and
wherein the instructions are pipelined in a manner which permits at least a given one of the threads to support multiple concurrent instruction pipelines.
2. The multithreaded processor of claim 1 wherein the instruction issuance sequence is determined using token triggered threading.
3. The multithreaded processor of claim 2 wherein processor supports N threads, and over a sequence of N consecutive processor clock cycles each of the N threads is permitted to issue instructions on only a corresponding one of the N consecutive processor clock cycles.
4. The multithreaded processor of claim 1 wherein each of the threads is permitted to issue two or more instructions on its corresponding one of the processor clock cycles.
5. The multithreaded processor of claim 1 wherein each of the threads is permitted to issue up to three instructions on its corresponding one of the processor clock cycles.
6. The multithreaded processor of claim 1 wherein the instructions are pipelined such that up to three instructions can be issued in a given clock cycle.
7. The multithreaded processor of claim 1 wherein the instructions are pipelined such that at least five separate instruction pipelines can be concurrently executing for different ones of the threads.
8. The multithreaded processor of claim 1 wherein the pipelined instructions comprise at least one of a load/store instruction, an arithmetic logic unit instruction, an integer multiplication instruction, a vector multiplication instruction, and a vector multiplication and reduction instruction.
9. The multithreaded processor of claim 8 wherein the arithmetic logic unit instruction has a pipeline comprising at least one wait stage between an instruction decode stage and a register file read stage thereof, the wait stage permitting the load/store instruction and the arithmetic logic instruction to share a set of register file read ports.
10. The multithreaded processor of claim 8 wherein the integer multiplication instruction has a pipeline comprising at least one wait stage between an instruction decode stage and a register file read stage thereof, the wait stage permitting the integer multiplication instruction and the arithmetic logic instruction to share a set of register file read ports.
11. The multithreaded processor of claim 8 wherein the vector multiplication instruction utilizes a different set of execution unit resources than the arithmetic logic unit and integer multiplication instructions, and can therefore execute concurrently with either of said instructions.
12. The multithreaded processor of claim 8 wherein the vector multiplication and reduction instruction is pipelined using a number of stages which is greater than a total number of threads of the processor.
13. The multithreaded processor of claim 8 wherein the vector multiplication and reduction instruction comprises a pipeline with at least eleven stages.
14. The multithreaded processor of claim 13 wherein the vector multiplication and reduction instruction comprises an instruction decode stage, a vector register file read stage, at least two multiply stages, at least two add stages, an accumulator read stage, a plurality of reduction stages, and an accumulator writeback stage.
15. The multithreaded processor of claim 14 wherein the accumulator read stage is combined with one of the add stages.
16. The multithreaded processor of claim 8 wherein pipelines for respective vector multiplication and reduction instructions are shifted relative to one another by a plurality of pipeline stages.
17. The multithreaded processor of claim 1 wherein the pipelined instructions comprise a vector multiplication and reduction instruction having a pipeline with a computational cycle which is longer than an issue cycle of the multithreaded processor.
18. The multithreaded processor of claim 1 wherein the processor comprises an integrated circuit.
19. An integrated circuit comprising:
a memory; and
a multithreaded processor coupled to the memory;
the multithreaded processor comprising:
a plurality of hardware thread units;
an instruction decoder coupled to the thread units for decoding instructions received therefrom; and
a plurality of execution units for executing the decoded instructions;
wherein the multithreaded processor is configured for controlling an instruction issuance sequence for a plurality of threads associated with respective ones of the hardware thread units;
wherein on a given processor clock cycle only a designated one of the threads is permitted to issue one or more instructions, the designated thread that is permitted to issue instructions varying over a plurality of clock cycles in accordance with the instruction issuance sequence; and
wherein the instructions are pipelined in a manner which permits at least a given one of the threads to support multiple concurrent instruction pipelines.
20. A method for use in a multithreaded processor comprising a plurality of hardware thread units associated with respective threads, an instruction decoder coupled to the thread units for decoding instructions received therefrom, and a plurality of execution units for executing the decoded instructions, the method comprising the steps of:
on a given processor clock cycle permitting only a designated one of the threads to issue one or more instructions;
varying the designated thread that is permitted to issue instructions over a plurality of clock cycles in accordance with an instruction issuance sequence; and
pipelining the instructions in a manner which permits at least a given one of the threads to support multiple concurrent instruction pipelines.
US11/096,917 2004-04-07 2005-04-01 Multithreaded processor with multiple concurrent pipelines per thread Active 2026-02-15 US8074051B2 (en)

Priority Applications (16)

Application Number Priority Date Filing Date Title
US11/096,917 US8074051B2 (en) 2004-04-07 2005-04-01 Multithreaded processor with multiple concurrent pipelines per thread
TW094111009A TW200604940A (en) 2004-04-07 2005-04-07 Multithreaded processor with multiple concurrent pipelines per thread
KR1020127022423A KR101236396B1 (en) 2004-04-07 2005-04-07 Multithreaded processor with multiple concurrent pipelines per thread
JP2007507460A JP2007533007A (en) 2004-04-07 2005-04-07 Multi-thread processor with multiple simultaneous pipelines per thread
EP05732166A EP1741033A4 (en) 2004-04-07 2005-04-07 Multithreaded processor with multiple concurrent pipelines per thread
EP11001890A EP2339457A3 (en) 2004-04-07 2005-04-07 Multithreaded processor with register file
KR1020127022422A KR101295569B1 (en) 2004-04-07 2005-04-07 Multithreaded processor with multiple concurrent pipelines per thread
EP11001889A EP2339456A3 (en) 2004-04-07 2005-04-07 Multithreaded processor with thread switch triggering
KR1020067022996A KR101253549B1 (en) 2004-04-07 2005-04-07 Multithreaded processor with multiple concurrent pipelines per thread
EP11001888A EP2339455A3 (en) 2004-04-07 2005-04-07 Multithreaded processor with interleaved instruction pipelines
KR1020127022421A KR101303119B1 (en) 2004-04-07 2005-04-07 Multithreaded processor with multiple concurrent pipelines per thread
PCT/US2005/011614 WO2005101221A1 (en) 2004-04-07 2005-04-07 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,893 US8892849B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,867 US8959315B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,912 US8918627B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US13/282,800 US8762688B2 (en) 2004-04-07 2011-10-27 Multithreaded processor with multiple concurrent pipelines per thread

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US56019904P 2004-04-07 2004-04-07
US11/096,917 US8074051B2 (en) 2004-04-07 2005-04-01 Multithreaded processor with multiple concurrent pipelines per thread

Related Parent Applications (1)

Application Number Title Priority Date Filing Date
US56019904P Continuation 2004-04-07 2004-04-07

Related Child Applications (4)

Application Number Title Priority Date Filing Date
US12/579,893 Continuation US8892849B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,912 Continuation US8918627B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,867 Continuation US8959315B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US13/282,800 Continuation US8762688B2 (en) 2004-04-07 2011-10-27 Multithreaded processor with multiple concurrent pipelines per thread

Publications (2)

Publication Number Publication Date
US20060095729A1 true US20060095729A1 (en) 2006-05-04
US8074051B2 US8074051B2 (en) 2011-12-06

Family

ID=35150180

Family Applications (5)

Application Number Title Priority Date Filing Date
US11/096,917 Active 2026-02-15 US8074051B2 (en) 2004-04-07 2005-04-01 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,893 Active US8892849B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,912 Active US8918627B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,867 Active US8959315B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US13/282,800 Active 2025-07-29 US8762688B2 (en) 2004-04-07 2011-10-27 Multithreaded processor with multiple concurrent pipelines per thread

Family Applications After (4)

Application Number Title Priority Date Filing Date
US12/579,893 Active US8892849B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,912 Active US8918627B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US12/579,867 Active US8959315B2 (en) 2004-04-07 2009-10-15 Multithreaded processor with multiple concurrent pipelines per thread
US13/282,800 Active 2025-07-29 US8762688B2 (en) 2004-04-07 2011-10-27 Multithreaded processor with multiple concurrent pipelines per thread

Country Status (6)

Country Link
US (5) US8074051B2 (en)
EP (4) EP1741033A4 (en)
JP (1) JP2007533007A (en)
KR (4) KR101303119B1 (en)
TW (1) TW200604940A (en)
WO (1) WO2005101221A1 (en)

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060227966A1 (en) * 2005-04-08 2006-10-12 Icera Inc. (Delaware Corporation) Data access and permute unit
US20070150624A1 (en) * 2005-11-04 2007-06-28 Sun Microsystems, Inc. Pipelined I/O execution
US20090193279A1 (en) * 2008-01-30 2009-07-30 Sandbridge Technologies, Inc. Method for enabling multi-processor synchronization
US20090235032A1 (en) * 2008-03-13 2009-09-17 Sandbridge Technologies, Inc. Method for achieving power savings by disabling a valid array
US20090276432A1 (en) * 2004-11-17 2009-11-05 Erdem Hokenek Data file storing multiple data types with controlled data access
US20100031007A1 (en) * 2008-02-18 2010-02-04 Sandbridge Technologies, Inc. Method to accelerate null-terminated string operations
US20100115527A1 (en) * 2006-11-10 2010-05-06 Sandbridge Technologies, Inc. Method and system for parallelization of pipelined computations
US20100122068A1 (en) * 2004-04-07 2010-05-13 Erdem Hokenek Multithreaded processor with multiple concurrent pipelines per thread
CN101290589B (en) * 2007-12-27 2010-06-16 华为技术有限公司 Parallel instruction operation method and device
US20100241834A1 (en) * 2007-11-05 2010-09-23 Sandbridge Technologies, Inc. Method of encoding using instruction field overloading
US20100293210A1 (en) * 2006-09-26 2010-11-18 Sandbridge Technologies, Inc. Software implementation of matrix inversion in a wireless communication system
US8732382B2 (en) 2008-08-06 2014-05-20 Qualcomm Incorporated Haltable and restartable DMA engine
US20140351568A1 (en) * 2011-11-15 2014-11-27 Wuxi Dsp Technologies Inc. Opportunistic multi-thread method and processor
CN104460934A (en) * 2013-09-13 2015-03-25 华为终端有限公司 Scheduling method and device for multi central processing units (CPU)
WO2015119953A1 (en) * 2014-02-06 2015-08-13 Optimum Semiconductor Technologies, Inc. Opportunity multithreading in a multithreaded processor with instruction chaining capability
KR20160118329A (en) * 2014-02-06 2016-10-11 옵티멈 세미컨덕터 테크놀로지스 인코포레이티드 Deterministic and opportunistic multithreading
US9766894B2 (en) 2014-02-06 2017-09-19 Optimum Semiconductor Technologies, Inc. Method and apparatus for enabling a processor to generate pipeline control signals
US20180113709A1 (en) * 2016-10-21 2018-04-26 Advanced Micro Devices, Inc. Method and system for performing low power and low latency multi-precision computation
US20210064365A1 (en) * 2019-08-29 2021-03-04 International Business Machines Corporation Instruction handling for accumulation of register results in a microprocessor
US11210098B2 (en) * 2013-07-15 2021-12-28 Texas Instruments Incorporated Variable latency instructions

Families Citing this family (26)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8195922B2 (en) 2005-03-18 2012-06-05 Marvell World Trade, Ltd. System for dynamically allocating processing time to multiple threads
US8099448B2 (en) 2005-11-02 2012-01-17 Qualcomm Incorporated Arithmetic logic and shifting device for use in a processor
US20080071851A1 (en) * 2006-09-20 2008-03-20 Ronen Zohar Instruction and logic for performing a dot-product operation
KR100888675B1 (en) * 2007-02-15 2009-03-13 고려대학교 산학협력단 Extended processor for excuting multi-threads in embeded core and Method for excuting multi-threads in embeded core
KR101041777B1 (en) * 2009-06-08 2011-06-17 엠텍비젼 주식회사 Method for processing multi thread and Processor device using the same
US20110107061A1 (en) * 2009-10-30 2011-05-05 Warren David A Performance of first and second macros while data is moving through hardware pipeline
US9325352B2 (en) * 2009-12-30 2016-04-26 Intel Deutschland Gmbh Control architectures for RF transceivers
JP5533330B2 (en) * 2010-06-23 2014-06-25 富士ゼロックス株式会社 Data processing device
JP2012048399A (en) * 2010-08-25 2012-03-08 Panasonic Corp Thread arbitration system, processor, video recording and reproducing device, and thread arbitration method
WO2013080289A1 (en) * 2011-11-28 2013-06-06 富士通株式会社 Signal processing device and signal processing method
KR101869325B1 (en) * 2011-12-13 2018-06-21 한국전자통신연구원 Core allocation apparatus in different multi-core
JP5819184B2 (en) * 2011-12-28 2015-11-18 富士通株式会社 Information processing apparatus and information processing apparatus control method
JP5751181B2 (en) * 2012-01-24 2015-07-22 富士通セミコンダクター株式会社 Instruction control circuit, processor, and instruction control method
US9588766B2 (en) * 2012-09-28 2017-03-07 Intel Corporation Accelerated interlane vector reduction instructions
US9430384B2 (en) * 2013-03-31 2016-08-30 Intel Corporation Instructions and logic to provide advanced paging capabilities for secure enclave page caches
US9558002B2 (en) * 2014-09-30 2017-01-31 Imagination Techologies Limited Variable length execution pipeline
KR102332523B1 (en) 2014-12-24 2021-11-29 삼성전자주식회사 Apparatus and method for execution processing
KR20160108754A (en) 2015-03-06 2016-09-20 주식회사 에이디칩스 Method for processing unconditional branch instruction in processor with pipeline
US9870340B2 (en) * 2015-03-30 2018-01-16 International Business Machines Corporation Multithreading in vector processors
KR101856487B1 (en) 2017-03-03 2018-06-19 주식회사 티맥스데이터 Computing device for processing parsing
CN108628639B (en) 2017-03-21 2021-02-12 华为技术有限公司 Processor and instruction scheduling method
US20200310799A1 (en) * 2019-03-27 2020-10-01 Mediatek Inc. Compiler-Allocated Special Registers That Resolve Data Hazards With Reduced Hardware Complexity
US11294672B2 (en) 2019-08-22 2022-04-05 Apple Inc. Routing circuitry for permutation of single-instruction multiple-data operands
US11256518B2 (en) 2019-10-09 2022-02-22 Apple Inc. Datapath circuitry for math operations using SIMD pipelines
US11126439B2 (en) 2019-11-15 2021-09-21 Apple Inc. SIMD operand permutation with selection from among multiple registers
US11855831B1 (en) 2022-06-10 2023-12-26 T-Mobile Usa, Inc. Enabling an operator to resolve an issue associated with a 5G wireless telecommunication network using AR glasses

Citations (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4001692A (en) * 1975-07-07 1977-01-04 Barry Research Corporation Time diversity data transmission apparatus
US4706211A (en) * 1983-09-22 1987-11-10 Sony Corporation Digital multiplying circuit
US4769779A (en) * 1985-12-16 1988-09-06 Texas Instruments Incorporated Systolic complex multiplier
US5181184A (en) * 1989-06-20 1993-01-19 Samsung Electronics Co., Ltd. Apparatus for multiplying real-time 2's complement code in a digital signal processing system and a method for the same
US5404469A (en) * 1992-02-25 1995-04-04 Industrial Technology Research Institute Multi-threaded microprocessor architecture utilizing static interleaving
US5613114A (en) * 1994-04-15 1997-03-18 Apple Computer, Inc System and method for custom context switching
US5864703A (en) * 1997-10-09 1999-01-26 Mips Technologies, Inc. Method for providing extended precision in SIMD vector arithmetic operations
US5889689A (en) * 1997-09-08 1999-03-30 Lucent Technologies Inc. Hierarchical carry-select, three-input saturation
US5907702A (en) * 1997-03-28 1999-05-25 International Business Machines Corporation Method and apparatus for decreasing thread switch latency in a multithread processor
US5949996A (en) * 1994-01-06 1999-09-07 Fujitsu Limited Processor having a variable number of stages in a pipeline
US5958041A (en) * 1997-06-26 1999-09-28 Sun Microsystems, Inc. Latency prediction in a pipelined microarchitecture
US5983256A (en) * 1995-08-31 1999-11-09 Intel Corporation Apparatus for performing multiply-add operations on packed data
US6078941A (en) * 1996-11-18 2000-06-20 Samsung Electronics Co., Ltd. Computational structure having multiple stages wherein each stage includes a pair of adders and a multiplexing circuit capable of operating in parallel
US6295600B1 (en) * 1996-07-01 2001-09-25 Sun Microsystems, Inc. Thread switch on blocked load or store using instruction thread field
US6377619B1 (en) * 1997-09-26 2002-04-23 Agere Systems Guardian Corp. Filter structure and method
US6530014B2 (en) * 1997-09-08 2003-03-04 Agere Systems Inc. Near-orthogonal dual-MAC instruction set architecture with minimal encoding bits
US6557022B1 (en) * 2000-02-26 2003-04-29 Qualcomm, Incorporated Digital signal processor with coupled multiply-accumulate units
US20030120901A1 (en) * 2001-12-20 2003-06-26 Erdem Hokenek Multithreaded processor with efficient processing for convergence device applications
US6687724B1 (en) * 1999-05-07 2004-02-03 Sony Corporation Information processor
US6697935B1 (en) * 1997-10-23 2004-02-24 International Business Machines Corporation Method and apparatus for selecting thread switch events in a multithreaded processor
US20040073772A1 (en) * 2002-10-11 2004-04-15 Erdem Hokenek Method and apparatus for thread-based memory access in a multithreaded processor
US20040073779A1 (en) * 2002-10-11 2004-04-15 Erdem Hokenek Method and apparatus for register file port reduction in a multithreaded processor
US6842848B2 (en) * 2002-10-11 2005-01-11 Sandbridge Technologies, Inc. Method and apparatus for token triggered multithreading
US6973471B2 (en) * 2002-02-22 2005-12-06 Freescale Semiconductor, Inc. Method and apparatus for implementing signed multiplication of operands having differing bit widths without sign extension of the multiplicand
US7593978B2 (en) * 2003-05-09 2009-09-22 Sandbridge Technologies, Inc. Processor reduction unit for accumulation of multiple operands with or without saturation

Family Cites Families (36)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5241635A (en) 1988-11-18 1993-08-31 Massachusetts Institute Of Technology Tagged token data processing system with operand matching in activation frames
US5812811A (en) * 1995-02-03 1998-09-22 International Business Machines Corporation Executing speculative parallel instructions threads with forking and inter-thread communication
JP2798121B2 (en) * 1995-08-04 1998-09-17 日本電気株式会社 Data processing device
JP3510729B2 (en) 1996-02-29 2004-03-29 三洋電機株式会社 Instruction execution method and instruction execution device
US6088788A (en) 1996-12-27 2000-07-11 International Business Machines Corporation Background completion of instruction and associated fetch request in a multithread processor
US6073159A (en) * 1996-12-31 2000-06-06 Compaq Computer Corporation Thread properties attribute vector based thread selection in multithreading processor
US6212544B1 (en) * 1997-10-23 2001-04-03 International Business Machines Corporation Altering thread priorities in a multithreaded processor
US6161166A (en) 1997-11-10 2000-12-12 International Business Machines Corporation Instruction cache for multithreaded processor
US5991785A (en) 1997-11-13 1999-11-23 Lucent Technologies Inc. Determining an extremum value and its index in an array using a dual-accumulation processor
US6092175A (en) * 1998-04-02 2000-07-18 University Of Washington Shared register storage mechanisms for multithreaded computer systems with out-of-order execution
US6182105B1 (en) 1998-08-27 2001-01-30 Lucent Technologies Inc. Multiple-operand addition with intermediate saturation
US6606704B1 (en) * 1999-08-31 2003-08-12 Intel Corporation Parallel multithreaded processor with plural microengines executing multiple threads each microengine having loadable microcode
US6526430B1 (en) 1999-10-04 2003-02-25 Texas Instruments Incorporated Reconfigurable SIMD coprocessor architecture for sum of absolute differences and symmetric filtering (scalable MAC engine for image processing)
US7120783B2 (en) 1999-12-22 2006-10-10 Ubicom, Inc. System and method for reading and writing a thread state in a multithreaded central processing unit
US6359579B1 (en) 2000-02-17 2002-03-19 Advanced Micro Devices, Inc. Digital logic correction circuit for a pipeline analog to digital (A/D) converter
US6694425B1 (en) 2000-05-04 2004-02-17 International Business Machines Corporation Selective flush of shared and other pipeline stages in a multithread processor
US6898694B2 (en) 2001-06-28 2005-05-24 Intel Corporation High instruction fetch bandwidth in multithread processor using temporary instruction cache to deliver portion of cache line in subsequent clock cycle
US7487505B2 (en) * 2001-08-27 2009-02-03 Intel Corporation Multithreaded microprocessor with register allocation based on number of active threads
US6912623B2 (en) 2002-06-04 2005-06-28 Sandbridge Technologies, Inc. Method and apparatus for multithreaded cache with simplified implementation of cache replacement policy
US6990557B2 (en) 2002-06-04 2006-01-24 Sandbridge Technologies, Inc. Method and apparatus for multithreaded cache with cache eviction based on thread identifier
GB2389433B (en) 2002-06-08 2005-08-31 Motorola Inc Bit exactness support in dual-mac architecture
US6971103B2 (en) 2002-10-15 2005-11-29 Sandbridge Technologies, Inc. Inter-thread communications using shared interrupt register
US7428567B2 (en) 2003-07-23 2008-09-23 Sandbridge Technologies, Inc. Arithmetic unit for addition or subtraction with preliminary saturation detection
US7251737B2 (en) 2003-10-31 2007-07-31 Sandbridge Technologies, Inc. Convergence device with dynamic program throttling that replaces noncritical programs with alternate capacity programs based on power indicator
US7360064B1 (en) * 2003-12-10 2008-04-15 Cisco Technology, Inc. Thread interleaving in a multithreaded embedded processor
US7873815B2 (en) 2004-03-04 2011-01-18 Qualcomm Incorporated Digital signal processors with configurable dual-MAC and dual-ALU
US7797363B2 (en) 2004-04-07 2010-09-14 Sandbridge Technologies, Inc. Processor having parallel vector multiply and reduce operations with sequential semantics
US7475222B2 (en) 2004-04-07 2009-01-06 Sandbridge Technologies, Inc. Multi-threaded processor having compound instruction and operation formats
US8074051B2 (en) 2004-04-07 2011-12-06 Aspen Acquisition Corporation Multithreaded processor with multiple concurrent pipelines per thread
TW200625097A (en) 2004-11-17 2006-07-16 Sandbridge Technologies Inc Data file storing multiple date types with controlled data access
US9110726B2 (en) 2006-11-10 2015-08-18 Qualcomm Incorporated Method and system for parallelization of pipelined computations
US20100299319A1 (en) 2007-08-31 2010-11-25 Sandbridge Technologies, Inc. Method, apparatus, and architecture for automated interaction between subscribers and entities
US20100241834A1 (en) 2007-11-05 2010-09-23 Sandbridge Technologies, Inc. Method of encoding using instruction field overloading
WO2009097444A1 (en) 2008-01-30 2009-08-06 Sandbridge Technologies, Inc. Method for enabling multi-processor synchronization
KR20100126690A (en) 2008-02-18 2010-12-02 샌드브리지 테크놀로지스, 인코포레이티드 Method to accelerate null-terminated string operations
US8762641B2 (en) 2008-03-13 2014-06-24 Qualcomm Incorporated Method for achieving power savings by disabling a valid array

Patent Citations (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4001692A (en) * 1975-07-07 1977-01-04 Barry Research Corporation Time diversity data transmission apparatus
US4706211A (en) * 1983-09-22 1987-11-10 Sony Corporation Digital multiplying circuit
US4769779A (en) * 1985-12-16 1988-09-06 Texas Instruments Incorporated Systolic complex multiplier
US5181184A (en) * 1989-06-20 1993-01-19 Samsung Electronics Co., Ltd. Apparatus for multiplying real-time 2's complement code in a digital signal processing system and a method for the same
US5404469A (en) * 1992-02-25 1995-04-04 Industrial Technology Research Institute Multi-threaded microprocessor architecture utilizing static interleaving
US5949996A (en) * 1994-01-06 1999-09-07 Fujitsu Limited Processor having a variable number of stages in a pipeline
US5613114A (en) * 1994-04-15 1997-03-18 Apple Computer, Inc System and method for custom context switching
US5983256A (en) * 1995-08-31 1999-11-09 Intel Corporation Apparatus for performing multiply-add operations on packed data
US6295600B1 (en) * 1996-07-01 2001-09-25 Sun Microsystems, Inc. Thread switch on blocked load or store using instruction thread field
US6078941A (en) * 1996-11-18 2000-06-20 Samsung Electronics Co., Ltd. Computational structure having multiple stages wherein each stage includes a pair of adders and a multiplexing circuit capable of operating in parallel
US5907702A (en) * 1997-03-28 1999-05-25 International Business Machines Corporation Method and apparatus for decreasing thread switch latency in a multithread processor
US5958041A (en) * 1997-06-26 1999-09-28 Sun Microsystems, Inc. Latency prediction in a pipelined microarchitecture
US6530014B2 (en) * 1997-09-08 2003-03-04 Agere Systems Inc. Near-orthogonal dual-MAC instruction set architecture with minimal encoding bits
US5889689A (en) * 1997-09-08 1999-03-30 Lucent Technologies Inc. Hierarchical carry-select, three-input saturation
US6377619B1 (en) * 1997-09-26 2002-04-23 Agere Systems Guardian Corp. Filter structure and method
US5864703A (en) * 1997-10-09 1999-01-26 Mips Technologies, Inc. Method for providing extended precision in SIMD vector arithmetic operations
US6697935B1 (en) * 1997-10-23 2004-02-24 International Business Machines Corporation Method and apparatus for selecting thread switch events in a multithreaded processor
US6687724B1 (en) * 1999-05-07 2004-02-03 Sony Corporation Information processor
US6557022B1 (en) * 2000-02-26 2003-04-29 Qualcomm, Incorporated Digital signal processor with coupled multiply-accumulate units
US20030120901A1 (en) * 2001-12-20 2003-06-26 Erdem Hokenek Multithreaded processor with efficient processing for convergence device applications
US6973471B2 (en) * 2002-02-22 2005-12-06 Freescale Semiconductor, Inc. Method and apparatus for implementing signed multiplication of operands having differing bit widths without sign extension of the multiplicand
US20040073772A1 (en) * 2002-10-11 2004-04-15 Erdem Hokenek Method and apparatus for thread-based memory access in a multithreaded processor
US20040073779A1 (en) * 2002-10-11 2004-04-15 Erdem Hokenek Method and apparatus for register file port reduction in a multithreaded processor
US6842848B2 (en) * 2002-10-11 2005-01-11 Sandbridge Technologies, Inc. Method and apparatus for token triggered multithreading
US7593978B2 (en) * 2003-05-09 2009-09-22 Sandbridge Technologies, Inc. Processor reduction unit for accumulation of multiple operands with or without saturation

Cited By (40)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8892849B2 (en) 2004-04-07 2014-11-18 Qualcomm Incorporated Multithreaded processor with multiple concurrent pipelines per thread
US20100199075A1 (en) * 2004-04-07 2010-08-05 Erdem Hokenek Multithreaded processor with multiple concurrent pipelines per thread
US8959315B2 (en) 2004-04-07 2015-02-17 Qualcomm Incorporated Multithreaded processor with multiple concurrent pipelines per thread
US20100122068A1 (en) * 2004-04-07 2010-05-13 Erdem Hokenek Multithreaded processor with multiple concurrent pipelines per thread
US8918627B2 (en) 2004-04-07 2014-12-23 Qualcomm Incorporated Multithreaded processor with multiple concurrent pipelines per thread
US20100199073A1 (en) * 2004-04-07 2010-08-05 Erdem Hokenek Multithreaded processor with multiple concurrent pipelines per thread
US20090276432A1 (en) * 2004-11-17 2009-11-05 Erdem Hokenek Data file storing multiple data types with controlled data access
US7933405B2 (en) * 2005-04-08 2011-04-26 Icera Inc. Data access and permute unit
US20060227966A1 (en) * 2005-04-08 2006-10-12 Icera Inc. (Delaware Corporation) Data access and permute unit
US7398329B2 (en) * 2005-11-04 2008-07-08 Sun Microsystems, Inc. Pipelined I/O execution
US20070150624A1 (en) * 2005-11-04 2007-06-28 Sun Microsystems, Inc. Pipelined I/O execution
US8819099B2 (en) 2006-09-26 2014-08-26 Qualcomm Incorporated Software implementation of matrix inversion in a wireless communication system
US20100293210A1 (en) * 2006-09-26 2010-11-18 Sandbridge Technologies, Inc. Software implementation of matrix inversion in a wireless communication system
US20100115527A1 (en) * 2006-11-10 2010-05-06 Sandbridge Technologies, Inc. Method and system for parallelization of pipelined computations
US9110726B2 (en) 2006-11-10 2015-08-18 Qualcomm Incorporated Method and system for parallelization of pipelined computations
US20100241834A1 (en) * 2007-11-05 2010-09-23 Sandbridge Technologies, Inc. Method of encoding using instruction field overloading
US20100262810A1 (en) * 2007-12-27 2010-10-14 Huawei Technologies Co., Ltd. Concurrent instruction operation method and device
CN101290589B (en) * 2007-12-27 2010-06-16 华为技术有限公司 Parallel instruction operation method and device
US8539188B2 (en) 2008-01-30 2013-09-17 Qualcomm Incorporated Method for enabling multi-processor synchronization
US20090193279A1 (en) * 2008-01-30 2009-07-30 Sandbridge Technologies, Inc. Method for enabling multi-processor synchronization
US20100031007A1 (en) * 2008-02-18 2010-02-04 Sandbridge Technologies, Inc. Method to accelerate null-terminated string operations
US8762641B2 (en) 2008-03-13 2014-06-24 Qualcomm Incorporated Method for achieving power savings by disabling a valid array
US20090235032A1 (en) * 2008-03-13 2009-09-17 Sandbridge Technologies, Inc. Method for achieving power savings by disabling a valid array
US8732382B2 (en) 2008-08-06 2014-05-20 Qualcomm Incorporated Haltable and restartable DMA engine
US20140351568A1 (en) * 2011-11-15 2014-11-27 Wuxi Dsp Technologies Inc. Opportunistic multi-thread method and processor
US9740498B2 (en) * 2011-11-15 2017-08-22 Wuxi Dsp Technologies Inc. Opportunistic multi-thread method and processor
US11210098B2 (en) * 2013-07-15 2021-12-28 Texas Instruments Incorporated Variable latency instructions
CN104460934A (en) * 2013-09-13 2015-03-25 华为终端有限公司 Scheduling method and device for multi central processing units (CPU)
KR20160118329A (en) * 2014-02-06 2016-10-11 옵티멈 세미컨덕터 테크놀로지스 인코포레이티드 Deterministic and opportunistic multithreading
US9558000B2 (en) 2014-02-06 2017-01-31 Optimum Semiconductor Technologies, Inc. Multithreading using an ordered list of hardware contexts
KR20160110528A (en) * 2014-02-06 2016-09-21 옵티멈 세미컨덕터 테크놀로지스 인코포레이티드 Opportunity multithreading in a multithreaded processor with instruction chaining capability
US9766894B2 (en) 2014-02-06 2017-09-19 Optimum Semiconductor Technologies, Inc. Method and apparatus for enabling a processor to generate pipeline control signals
WO2015119953A1 (en) * 2014-02-06 2015-08-13 Optimum Semiconductor Technologies, Inc. Opportunity multithreading in a multithreaded processor with instruction chaining capability
KR102335194B1 (en) 2014-02-06 2021-12-02 옵티멈 세미컨덕터 테크놀로지스 인코포레이티드 Opportunity multithreading in a multithreaded processor with instruction chaining capability
KR102341785B1 (en) 2014-02-06 2021-12-20 옵티멈 세미컨덕터 테크놀로지스 인코포레이티드 Deterministic and opportunistic multithreading
US20180113709A1 (en) * 2016-10-21 2018-04-26 Advanced Micro Devices, Inc. Method and system for performing low power and low latency multi-precision computation
US11409536B2 (en) * 2016-10-21 2022-08-09 Advanced Micro Devices, Inc. Pairing SIMD lanes to perform double precision operations
US20210064365A1 (en) * 2019-08-29 2021-03-04 International Business Machines Corporation Instruction handling for accumulation of register results in a microprocessor
US11132198B2 (en) * 2019-08-29 2021-09-28 International Business Machines Corporation Instruction handling for accumulation of register results in a microprocessor
US11755325B2 (en) 2019-08-29 2023-09-12 International Business Machines Corporation Instruction handling for accumulation of register results in a microprocessor

Also Published As

Publication number Publication date
US20100122068A1 (en) 2010-05-13
US20100199075A1 (en) 2010-08-05
EP2339456A3 (en) 2011-08-24
EP1741033A4 (en) 2008-07-09
EP2339457A2 (en) 2011-06-29
EP1741033A1 (en) 2007-01-10
EP2339456A2 (en) 2011-06-29
EP2339455A3 (en) 2011-08-24
KR101295569B1 (en) 2013-08-12
TW200604940A (en) 2006-02-01
WO2005101221A1 (en) 2005-10-27
KR20070011434A (en) 2007-01-24
US8959315B2 (en) 2015-02-17
US8892849B2 (en) 2014-11-18
JP2007533007A (en) 2007-11-15
KR20120114380A (en) 2012-10-16
KR20120115553A (en) 2012-10-18
US20120096243A1 (en) 2012-04-19
KR20120114379A (en) 2012-10-16
EP2339455A2 (en) 2011-06-29
US8918627B2 (en) 2014-12-23
KR101303119B1 (en) 2013-09-06
KR101253549B1 (en) 2013-04-11
KR101236396B1 (en) 2013-02-22
US8762688B2 (en) 2014-06-24
EP2339457A3 (en) 2011-08-31
US20100199073A1 (en) 2010-08-05
US8074051B2 (en) 2011-12-06

Similar Documents

Publication Publication Date Title
US8074051B2 (en) Multithreaded processor with multiple concurrent pipelines per thread
US7475222B2 (en) Multi-threaded processor having compound instruction and operation formats
US6842848B2 (en) Method and apparatus for token triggered multithreading
US7593978B2 (en) Processor reduction unit for accumulation of multiple operands with or without saturation
US5764943A (en) Data path circuitry for processor having multiple instruction pipelines
CN101322111A (en) Multithreading processor with each threading having multiple concurrent assembly line
US6092184A (en) Parallel processing of pipelined instructions having register dependencies
US7793072B2 (en) Vector execution unit to process a vector instruction by executing a first operation on a first set of operands and a second operation on a second set of operands
US8200945B2 (en) Vector unit in a processor enabled to replicate data on a first portion of a data bus to primary and secondary registers
KR20070018066A (en) Processor having compound instruction and operation formats

Legal Events

Date Code Title Description
AS Assignment

Owner name: SANDBRIDGE TECHNOLOGIES, INC., NEW YORK

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:GLOSSNER, C. JOHN;MOUDGILL, MAYAN;SCHULTE, MICHAEL J.;AND OTHERS;SIGNING DATES FROM 20050929 TO 20050930;REEL/FRAME:016950/0482

Owner name: SANDBRIDGE TECHNOLOGIES, INC., NEW YORK

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:GLOSSNER, C. JOHN;MOUDGILL, MAYAN;SCHULTE, MICHAEL J.;AND OTHERS;REEL/FRAME:016950/0482;SIGNING DATES FROM 20050929 TO 20050930

AS Assignment

Owner name: ASPEN ACQUISITION CORPORATION, CALIFORNIA

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:SANDBRIDGE TECHNOLOGIES, INC.;REEL/FRAME:025094/0793

Effective date: 20100910

STCF Information on status: patent grant

Free format text: PATENTED CASE

CC Certificate of correction
AS Assignment

Owner name: QUALCOMM INCORPORATED, CALIFORNIA

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ASPEN ACQUISITION CORPORATION;REEL/FRAME:029377/0700

Effective date: 20120927

FEPP Fee payment procedure

Free format text: PAYOR NUMBER ASSIGNED (ORIGINAL EVENT CODE: ASPN); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY

FPAY Fee payment

Year of fee payment: 4

MAFP Maintenance fee payment

Free format text: PAYMENT OF MAINTENANCE FEE, 8TH YEAR, LARGE ENTITY (ORIGINAL EVENT CODE: M1552); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY

Year of fee payment: 8

MAFP Maintenance fee payment

Free format text: PAYMENT OF MAINTENANCE FEE, 12TH YEAR, LARGE ENTITY (ORIGINAL EVENT CODE: M1553); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY

Year of fee payment: 12